pub fn qsa_gather(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
k_cache: DevicePtr,
v_cache: DevicePtr,
block_table: DevicePtr,
sel: DevicePtr,
k_out: DevicePtr,
v_out: DevicePtr,
n_sel: u32,
block_size: u32,
nkv: u32,
hd: u32,
stream: u64,
) -> Result<()>Expand description
Pack the selected tokens’ K/V rows into contiguous NHD scratch.